Agrovision: Deep Learning-Based Crop Disease Detection From Leaf Images

Abstract

This paper describes an organized approach that is used to identify and categorize crop diseases through deep learning and uses image data (taken on crop leaves). The paper uses an investigation to examine how convolutional neural networks (CNNs) and especially the transfer learning practice can be used to improve disease detection across various crop species. The model was able to be precise in identifying disease patterns through training on crop-based data and subsequently displaying a high accuracy in classification with a fairly small quantity of training samples. Rotation, flipping, and zooming augmentation methods are strategically used to reinforce the model performance and the ability to handle more datasets with consistency and therefore counter the effects of small-scale and unbalanced datasets. The paper also explores real-time location of disease using state-of-the-art object detection models such as YOLO so that infected areas in leaf images can be accurately identified and annotated to have fine-grained management of the disease. Such a thorough comparison of different deep learning architectures and the measure of their performances is implemented to conclude which of the deep learning architectures is the most viable in practical implementation in precision agriculture. As the findings show, the CNN-based models, especially those with the use of the transfer learning, are more accurate and efficient when it comes to both the predictive qualities and the efficiency of the models. These results point to the scalability and feasibility of deep learning to detect and then interfere early in the disease and to achieve lasting crop management practices. Finally, the study contributes to the body of agricultural automation research which enables farmers to seek proactive approaches to crop protection.